Bidirectional f-Divergence-Based Deep Generative Method for Imputing Missing Values in Time-Series Data.

Area of Science:

• Machine Learning
• Statistics
• Data Science

Background:

• Imputing missing values in high-dimensional time-series data is a significant challenge.
• Existing methods often struggle with high missing rates and reduced accuracy.
• Deep learning approaches have shown promise but require further refinement.

Purpose of the Study:

• To present a novel imputation network, tf-BiGAIN, for high-dimensional time-series data.
• To address limitations of existing methods, particularly concerning high missing rates.
• To improve accuracy and robustness in time-series imputation.

Main Methods:

• Developed a novel f-divergence-based bidirectional generative adversarial imputation network (tf-BiGAIN).
• Utilized bidirectional modified gated recurrent units for capturing temporal dependencies.
• Employed f-divergence as an objective function for model optimization without distributional assumptions.

Main Results:

• tf-BiGAIN demonstrated superior performance on two real-world time-series datasets.
• The method outperformed existing imputation techniques in terms of accuracy and robustness.
• The f-divergence framework and bidirectional architecture enhanced imputation capabilities.

Conclusions:

• tf-BiGAIN offers a flexible and adaptable solution for time-series data imputation.
• The bidirectional approach effectively leverages past and future temporal information.
• This novel network provides a robust and accurate method for handling missing data in complex time-series scenarios.